Ozone learns how your codebase works, then flags exploitable bugs on the pull request, before they merge.
2-minute install · $100 free credit
The transfer to msg.sender executes before balances is decremented, so a contract receiving it can re-enter emergencyWithdraw and drain the vault in one transaction. Move the state update above the call, or apply the existing nonReentrant modifier.
- (bool ok, ) = msg.sender.call{value: amount}("");+ balances[msg.sender] -= amount;
Illustrative data. Select a finding to open it.
#1on OpenAI's exploit benchmark, EVMBench
Your team writes code faster than any human can read it, and the exploitable bug is never the one the tests cover. Ozone reads every line at the speed the agents write it.
OZ_MODEL
Generic scanners pattern-match line by line. Ozone builds a working model of your system: how requests flow, which functions move funds, where untrusted input crosses a trust boundary. That model updates itself as your code evolves, and every review is judged against it rather than against a rule list.
OZ_REVIEW
Not a linter with a language model bolted on. The same multi-agent engine, pointed at the diff.
Triggers
Point Ozone at a repository and it reviews each pull request as it opens. Schedules cover the rest of the codebase on their own cadence.
Runs
Every run keeps its transcript: which files the agent read, which callers it followed, what it reasoned before it flagged anything. You can review the reviewer.
Findings
Each finding arrives classified with the concrete impact spelled out, not a generic rule ID and a link to a CWE page. Triage it as fixed, ignored, or a false positive.
Fixes
- msg.sender.call{value: amt}("");+ balances[msg.sender] -= amt;
Findings carry the suggested change as a diff, in the file and at the line where the problem was introduced.
Pull requests
Ozone runs as a check and comments on the pull request. It never pushes code. Findings land before the merge button, which is the only moment a fix is cheap.
Knowledge
Add specs, invariants, threat models or a past postmortem. They are loaded into the workspace so reviews know what your system is supposed to do.
OZ_BOUNDARY
Real attacks rarely respect the boundary. A leaked backend key drains a treasury. A compromised API feeds a bad price to a contract that trusted it. Ozone models both sides as one system and catches the issues that live in between. If you have no smart contracts at all, it works as continuous security review for a pure Web2 stack.
OZ_TRUST
Each review clones your repository into its own isolated sandbox. Nothing is shared between runs or between customers, and the sandbox is destroyed the moment the run ends.
The clone lives only inside that sandbox. No copy of your repository sits on our servers. What Ozone keeps is the review itself: the findings and the run transcript, the record of what the agent read and did.
Your code is never used to train AI models, and product analytics never include repository content. Ozone never pushes code, it writes only pull request comments and checks.
You choose which repositories the GitHub App can reach, and can revoke access at any time. Read the details
OZ_PRICING
No subscription needed to start. Pay only for what you run.
StandardStart free, pay for what you run.
$0 / month
$100 free credit included
ProFor teams running agents at scale.
$200 / month
Connect your repos and Ozone starts building its model. The first review runs on your next pull request.